Skip to content

LlamaIndex (Python)

RAG-first agent framework — data connectors, vector indexing, query engines, and workflow-based agents.

GroupData & RAG
Current versionv0.14.26
Bonds withPy

The data framework for LLM apps. LlamaIndex turns unstructured data into agent-queryable knowledge — document loaders, vector indices, query engines, tools, and event-driven workflows with prebuilt agents.

Install

Terminal window
pip install llama-index==0.14.26

Version

llama-index 0.14.26 (core and bundle share a version) · llama-index-workflows 2.26.1 · Python 3.10+ · checked 2026-11-19

Best for

RAG systems, document Q&A, enterprise knowledge agents.

Numbers are the comprehensive guide’s section numbers; unnumbered steps are separate pages.

Patch releases since the guide was last revised (from the upstream changelog):

  • 0.14.25 (September 2026) — restored streaming for the compact and refine response modes; native async for StructuredLLMRerank; removed the deprecated IPEX integrations; security dependency updates.
  • 0.14.24 (August 2026) — AgentWorkflow now honours an agent’s output_cls / structured_output_fn; Memory accepts any AsyncDBChatStore; async LLMRerank; many ingestion and retrieval fixes.
  • 0.14.23 (June 2026) — multimodal query engines and synthesis; a tool-calling mock LLM for tests; initial_state is deep-copied per run.
  • 0.14.22 (May 2026) — multimodal synthesis groundwork; ingestion cache and instrumentation fixes.

No breaking API changes in this range. The last breaking releases were 0.13.0 (legacy agents and QueryPipeline removed) and 0.14.0 (moved to llama-index-workflows 2.x).

import asyncio
from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.core.agent.workflow import FunctionAgent
from llama_index.core.tools import QueryEngineTool
from llama_index.llms.openai import OpenAI
documents = SimpleDirectoryReader("data").load_data()
index = VectorStoreIndex.from_documents(documents)
tool = QueryEngineTool.from_defaults(
query_engine=index.as_query_engine(),
name="knowledge_base",
description="Answer questions about the documents",
)
agent = FunctionAgent(tools=[tool], llm=OpenAI(model="gpt-6-luna"))
async def main():
response = await agent.run(user_msg="What are the key findings?")
print(response)
asyncio.run(main())

Ready? Start Zero → Hero Step 1 →


DateVersionChanges
2026-10-240.14.25 / workflows 2.26.1Audit against an installed release. New: response modes, chat engine modes, filters, chunking and metadata extraction, sentence-window and auto-merging retrieval, HyDE, citations (comprehensive guide); span and event reference, token budgets and local tracing (observability); chat-engine route, client and tests (streaming server); cited-answer and text-to-SQL recipes. Corrected: the observability page said agents emit AgentToolCallEvent (they do not) and that token counts are only tokenizer estimates (LLM counts use the provider’s usage when reported); the ReAct and router diagrams.
2026-10-240.14.25 / workflows 2.26.1Added the Workflows page; the Zero → Hero step for custom workflows now points to it.
2026-10-210.14.25Zero → Hero cards now carry the comprehensive guide’s section numbers; links added for the new property graph, multimodal, LlamaParse, serving and evaluation-dataset sections; production guide rewritten.
2026-10-210.14.25Version now read from the site’s version registry. Minimal example rewritten for FunctionAgent (the previous one used the removed ReActAgent.from_tools / .chat() API). llama-agents references replaced with AgentWorkflow and LlamaAgents/llamactl. Comprehensive guide, advanced pages and recipes re-verified against an installed copy.
2026-05-190.14.22Version bump.
2026-04-210.14.20Index redesigned with Zero → Hero + Jump-to-topic grid.
November 20250.14.8Initial documented version.

Note: TypeScript implementation lives at LlamaIndex (TypeScript) — that project has since been deprecated upstream.